Table 1 Review of statistical analysis methods [22,23,24,25,26,27]
Analysis method | Outcome type | Statistical assumptions | Advantages | Disadvantages |
|---|---|---|---|---|
Binary logistic regression (BLR) | Binary | • No assumptions made about explanatory variables | • Can adjust for covariates | • Large number of observations required |
Cox proportional hazards (CPH) | Binary | • Proportionality of hazards over time • Censoring of observations is unrelated to prognosis | • Can adjust for covariates | • If assumptions of the model not met then subsequent analyses and risk estimates will possibly be biased |
Chi-square (χ2) (CS) | Binary and ordered categorical | • Chi-Square – Total count is > 40 or total count is 20–40 and the expected value of each exposure-outcome category is > 5 | • Simple to implement | • Cannot adjust for covariates |
Cochran-Armitage trend test (CAT) | Ordered categorical | • Similar to the Chi-square test but it takes into account the ordering across categories | • Easy to interpret | • Cannot adjust for covariates |
Ordinal logistic regression (OLR) | Ordered categorical | • Response is ordinal • Proportionality of odds | • Can adjust for covariates | • If assumptions of the model not met then subsequent analyses and odds estimates will possibly be biased |
Mann-Whitney U test (MWU) | Ordered categorical | • Non-parametric test • Response is ordinal / continuous • Observations from both groups are independent of one another | • Easy to interpret | • Cannot adjust for covariates – there are extensions of this method, which allow for adjustment [28,29,30] |
Median test (MT) | Ordered categorical | • Non-parametric test • Considers the position of each observation relative to the overall median. | • Easy to interpret | • Cannot adjust for covariates • Inefficient (low power) to detect differences if sample size is large. |
t-test | Continuous (used on the ordered categorical) | • Response is continuous • Homogeneity of variances | • Easy to interpret | • Cannot adjust for covariates |
Multiple linear regression (MLR) | Continuous (used on the ordered categorical) | • Response is continuous • Linear relationship • Homogeneity of variances • No or little multicollinearity | • Can adjust for covariates | • Assumes linear relationship • Sensitive to outliers |
Win Ratio testWins/losses version (WR) | Combination of binary outcomes | • Responses for each outcome are binary • Accounts for clinical priorities of endpoints | • Prioritises the more major component of the outcome • Useful for composite outcomes • Extensions of this approach allow for covariate adjustment [31] • Easy to interpret | • New method • Doesn’t use the precise times from randomisation to event occurrence |
Bootstrapping (BS) | Ordered categorical | • None | • No assumptions made about the distribution of the data | • Cannot adjust for covariates • Computationally intensive • Doesn’t provide a meaningful point estimate |